Tran Hoang (Tommy) Soft

Senior Full-Stack Engineer | AI-Powered Financial Systems

Building production-grade intelligent platforms that combine real-time AI, financial data pipelines, and scalable cloud architecture.

---

🚀 Featured Project: 2winai

AI Stock Risk & Watchlist Platform — Full-stack, production-deployed, built and maintained by a single senior engineer.

Full-stack AI platform for financial market analysis, intelligent conversation, and automated risk & alert management.
Supports Web (Flask) and Mobile (React Native) with a unified backend deployed on Google Cloud Run.

2winai is composed of 5 independent, self-contained systems — each operates autonomously with no dependency on the others:

---

🤖 System 1 — AI Chatbot Engine

A conversational AI interface for market queries and in-depth financial analysis.

• Natural language understanding with custom-trained Rasa NLU (intent classification + entity extraction)
• Two independent modes, each with its own dedicated system prompt:
  • Normal Chat — General financial Q&A
  • Expert Analysis — Deep-dive queries with enriched live market context
• Security validation applied on all user inputs before processing
• Multi-LLM backend with primary/fallback strategy (see LLM Strategy below)

---

⚠️ System 2 — Risk Analysis Engine

Automated, multi-dimensional risk assessment — runs independently, triggered on demand or via scheduled jobs.

• Analyzes multiple risk dimensions using LLM inference (API-based only — no model training)
• Background processing with real-time UI progress: each dimension result appears incrementally as it completes
• Designed to be modular; dimensions can be extended or reconfigured without affecting other systems

---

🔔 System 3 — Alert & Watchlist Agent

Autonomous background agent for continuous market monitoring — fully decoupled from Systems 1 and 2.

• Monitors user watchlists on a scheduled basis
• Detects significant trend changes and signal anomalies using LLM inference (API-based only — no model training)
• Delivers timely notifications via email and in-app alerts
• Operates entirely in the background with no user interaction required

---

🛠️ System 4 — Rasa Training Manager

An internal desktop/web tool for managing the full lifecycle of the custom Rasa NLU model — built, tested, and operated entirely in-house.

• Step-by-step guided training pipeline:
  1. Run Docker Desktop
  2. Start Rasa Server
  3. Test Rasa Connection
  4. Run Training (force retrain)
  5. Complete & Log
• Real-time Training Logs panel — live output streamed directly from the Rasa process
• Intent Testing — enter any question and instantly verify the classified intent against the live model
• Quick Commands reference panel — Docker and Rasa CLI commands always visible
• Status indicators — Docker Desktop / Rasa Server / Training Status tracked in real time
• Actions: Check Docker Status · Download Logs · Reset · Check Docker (ps) · Stop Docker Rasa

This system ensures the Rasa model can be reliably rebuilt, validated, and deployed by any team member — no CLI expertise required.

---

🔧 System 5 — Admin Prompt & Data Management

A full-featured internal admin panel for managing AI prompts, user settings, and operational data — the control center of the 2winai intelligence layer.

Modules:

Module	Description
📊 Prompt Dashboard	View and manage all prompt templates in use across the platform
➕ Create Prompt	Create new prompts with: Prompt ID · Category · Mode · Intent · Rasa Keywords · Description · Prompt Template
⚠️ Error Records	All reported issues and system errors are automatically saved to the database for tracking, audit, and resolution
⚙️ API Settings	Configure LLM API keys, model selection, fallback priority, and endpoint settings
👤 User View & Settings	Manage user accounts, quota limits, subscription tiers, and access permissions
🗄️ Data Management	Manage financial data sources, cache, and pipeline configurations

Prompt Creation Fields:

• Prompt ID — Unique identifier for each prompt template
• Category — Logical grouping (e.g., risk, chat, alert)
• Mode — Normal / Expert / Custom
• Intent — Mapped Rasa intent that triggers this prompt
• Rasa Keywords — Comma-separated keywords used in Rasa training data (e.g., EPS, P/E, revenue)
• Description — Human-readable explanation of what this prompt does
• Prompt Template — The actual LLM system prompt body

The tight integration of Rasa intents and prompt templates in this admin panel is what enables precise, intent-driven LLM responses — a core architectural differentiator of 2winai.

---

🐛 Issue Reporting → Database

All user-reported issues and system error events are captured, structured, and persisted to PostgreSQL (Cloud SQL):

• Users can report issues directly from the UI
• Each report is timestamped, categorized, and linked to user session context
• Errors from background agents (Risk Engine, Alert Agent) are also auto-logged
• The Error Records module in the Admin Panel provides full visibility and audit trail
• Enables proactive monitoring, triage, and product improvement — no issues get lost

---

🧠 Multi-LLM Strategy

Role	Models
Primary	One of: Grok, Gemini, or ChatGPT — selected based on reliability and performance
Fallback	The remaining models activate automatically when the primary is unavailable, slow, or degraded
Extended	OpenRouter used occasionally for additional model access

• The caller (chat, risk, alert) always knows which model is active — fallback is explicit and logged, not hidden
• All LLMs are consumed via API inference only — no fine-tuning, no training, no custom model development

---

🏗️ Architecture Overview

┌────────────────────────────────────────────────────────────────────────┐
│                          2winai Platform                               │
│              (Google Cloud Run + Docker + GitHub Actions CI/CD)        │
├──────────────┬──────────────┬──────────────┬──────────────┬────────────┤
│  System 1    │  System 2    │  System 3    │  System 4    │  System 5  │
│  AI Chatbot  │  Risk        │  Alert &     │  Rasa        │  Admin     │
│  Engine      │  Analysis    │  Watchlist   │  Training    │  Prompt &  │
│              │  Engine      │  Agent       │  Manager     │  Data Mgmt │
│              │              │              │              │            │
│ • Rasa NLU   │ • Multi-dim  │ • Scheduled  │ • Build &    │ • Prompt   │
│ • Normal /   │   Risk Score │   Agent      │   Retrain    │   CRUD     │
│   Expert     │ • Real-time  │ • Trend Det. │ • Live Logs  │ • Intent   │
│   mode       │   Progress   │ • Email +    │ • Intent     │   Mapping  │
│ • Prompts    │ • Background │   In-app     │   Testing    │ • Error DB │
│ • Security   │   Threading  │   Notify     │ • Docker Mgmt│ • User Mgmt│
└──────────────┴──────────────┴──────────────┴──────────────┴────────────┘
                        │            │             │
                        └────────────┴─────────────┘
                              (each system independent)
                                       │
                       ┌───────────────▼───────────────┐
                       │        Multi-LLM Layer         │
                       │  Primary: Grok / Gemini /      │
                       │  ChatGPT  +  Fallback chain    │
                       └───────────────┬───────────────┘
                                       │
                       ┌───────────────▼───────────────┐
                       │    Financial Data APIs         │
                       │    PostgreSQL (Cloud SQL)      │
                       │    Issue & Error DB (Cloud SQL)│
                       └───────────────────────────────┘

---

💼 Technical Expertise

• Backend & Architecture: Python, Flask, PostgreSQL, Background threading, Real-time streaming progress, Modular independent service design
• AI Layer: Rasa NLU (custom build, train, test pipeline), Multi-LLM orchestration, Prompt engineering, Intent-driven prompt routing, Input security validation
• Data & Integration: Multiple financial market APIs, Real-time data enrichment pipelines, Issue & error tracking persistence
• Cloud & DevOps: Google Cloud Run, GitHub Actions CI/CD, Docker, Cloud SQL
• Admin & Tooling: Internal admin panel (prompt CRUD, user management, error records, API config), Rasa Training Manager desktop tool
• Security & Scalability: Google OAuth, JWT, Rate limiting, Secure input pipeline, Usage quota enforcement
• Payment Integration: PayPal (global subscriptions & payments), Google Pay (Google-ecosystem billing), VNPay (Vietnam domestic payment gateway) — full payment lifecycle: checkout, webhook handling, subscription management, and quota enforcement

---

📈 Engineering Philosophy

I focus on simplicity, reliability, and clear user value.
Each system is independently deployable and maintainable — observable in production, designed to fail gracefully, and built to deliver visible results step-by-step.

The combination of custom Rasa NLU (for precise, controllable intent routing) with multi-LLM inference (for language generation) reflects a deliberate architectural choice: control where it matters, flexibility where it helps.

---

📫 Get in Touch

• Location: Vietnam
• Open to: Senior Software Engineer / AI Engineer / Full-Stack Fintech roles (remote or onsite)
• Email: tranhoangsoft@gmail.com

---

Last updated: May 2026

---

"Crafting intelligent systems that deliver clear value — one reliable component at a time."